They saw a 90 percent improvement in patients who had low prediction performance.
Across the world, roughly 65 million people have epilepsy. The neurological disorder makes these individuals uncertain of when they might experience a seizure, burdening everyday life and, in some cases, breeding social stigma.
But what if patients with epilepsy and their healthcare providers could predict when the next seizure might strike?
The technology exists to do just that, but it is limited to certain kinds of patients. To spur development of additional predictive tools, researchers from University of Melbourne in Australia ran a seizure prediction contest on the data science website Kaggle. In the end, they crowdsourced 10,000 algorithms from 646 developers and 478 teams, finding that the top performers produced a 90 percent improvement in seizure prediction for the most challenging patients, according to a new study.
>> READ: Targeting Diabetes with Big Data, Machine Learning, Real-Time Informatics
“Epilepsy is highly different among individuals,” said Levin Kuhlmann, Ph.D., who works for the Graeme Clarke Institute and St. Vincent’s Hospital, Melbourne. “Results showed different algorithms performed best for different patients, supporting the use of patient-specific algorithms and long-term monitoring.”
In 2016, participants designed their algorithms to differentiate between 10-mintue inter-seizures and pre-seizure data clips, according to the study. They used long-term electrical brain activity recordings, which researchers collected from humans who took part in clinical trials for the NeuroVista Seizure Advisory System, an implantable prediction tool.
Then the study authors evaluated the algorithms and tested the best of the batch on patients who had the lowest seizure prediction performance in earlier studies. The result: a 90 percent improvement from prior results, with a sensitivity increase 1.9 times greater than the original trial, according to the study.
“These results indicate that clinically relevant seizure prediction is possible in a wider range of patients than previously thought possible,” the authors wrote. “Moreover, different algorithms performed best for different patients.”
Capitalizing on their findings, the researchers created a website called EpilepsyEcoSystem, which they intend to foster algorithm and data sharing in support of stronger, more personalized seizure prediction. They hope that continued crowdsourcing from top data scientists will ultimately scale real-time predictive technologies.
“The hope is to make seizures less like earthquakes, which can strike without warning, and more like hurricanes, where you have enough advanced warning to seek safety,” Kuhlmann said.
Get the best insights in healthcare analytics directly to your inbox.
An Algorithm to Detect Low Blood Pressure During Surgery
A Curious, New Hardware Fix for Cybersecurity Vulnerabilities
How an Algorithm Helped Rat Poison Save the Life of a President